We undertook the analysis on diabetes in the Prevention of Suicide in Primary Care Elderly: Collaborative Trial (PROSPECT) cohort after identifying older persons with diabetes as a subgroup for whom the risk of death has been reported to be increased by coexisting depression (1). We were guided in this exploratory analysis by published criteria for performing and reporting subgroup analyses (2).
At no time did we use automated variable selection methods as Thombs and Ziegelstein (3) suggest. We agree that overfitting is an important problem that deserves careful consideration. We had a prespecified approach to identifying and including potential confounders because we knew that imbalances would be likely and adjustment with patient-level variables would be necessary given the practice-randomized design. Our final model met the recommended rule of 10–15 events per predictor (4). We looked at the univariate relationship between potential confounders and the outcome of interest using a higher α-level for rejecting the null hypothesis of no confounding. This approach has been shown via simulation studies to yield acceptable confounder selection performance (5). For example, because individuals in the intervention group with diabetes were older at baseline compared with individuals in usual care with diabetes (mean age 71 ± 8.5 versus 67 ± 6.8 years), there was a bias toward the null hypothesis (i.e., that there was no intervention effect on mortality). Nevertheless, adjusting only for age, the intent-to-treat hazard ratio and corresponding 95% CI for patients with diabetes was consistent with the reported result (age-adjusted hazard ratio 0.46 [95% CI 0.26–0.81]). Because of the imbalance regarding age with respect to diabetes groups, the age-adjusted estimates of treatment effect may be closer to the true treatment effect.
Practices randomized to the intervention condition had available a number of components, including educational sessions for primary care physicians, education of patients’ families, and a depression care manager who worked within the practice. The care manager implemented the intervention by reviewing patients’ depression status, medical history, and medication use and subsequently worked with the primary care physician to recommend treatment according to standard guidelines, including medication and psychotherapy (6). The analogy with the commentary cited that focused on the effects of psychotherapy among persons with cancer does not apply to a multicomponent intervention.
The study design underwent rigorous peer review by National Institutes of Health panels and the reviewers of the journals in which the results were published. Large-scale intervention studies carried out in primary care practice are limited; therefore, we need to make the most of the data we have from these sources and external data such as mortality data, following in the tradition of well-established and accepted studies such as Framingham and Women's Health Initiative and entailing extended follow-up sub-studies. Although the PROSPECT intervention was not specifically designed to test whether a depression management program improved survival, the rationale for studying survival of the cohort rested on the many studies showing an association between depression and increased mortality.
